forked from IBM/Multi-GNN
-
Notifications
You must be signed in to change notification settings - Fork 0
/
training.py
393 lines (331 loc) · 17.2 KB
/
training.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
import torch
import tqdm
from sklearn.metrics import f1_score
from train_util import AddEgoIds, extract_param, add_arange_ids, get_loaders, evaluate_homo, evaluate_hetero, save_model, load_model
from models import GINe, PNA, GATe, RGCN, GINe_FHE, Model_Wrapper
from torch_geometric.data import Data, HeteroData
from torch_geometric.nn import to_hetero, summary
from torch_geometric.utils import degree
from torch.utils.data import TensorDataset
from concrete.ml.torch.compile import compile_brevitas_qat_model
import numpy as np
import wandb
import logging
import time
import tempfile
from pathlib import Path
def compile_and_test(tr_loader, te_loader, inds, torch_model, tr_data, te_data, args, use_sim=True):
X_test = te_data.x
y_test = te_data.y
for batch in tqdm.tqdm(te_loader, disable=not args.tqdm):
batch = batch
break
#remove the unique edge id from the edge features, as it's no longer needed
batch.edge_attr = batch.edge_attr[:, 1:]
#wrapped_model = Model_Wrapper(torch_model)
wrapped_model = Model_Wrapper(torch_model, batch.x, batch.edge_index, batch.edge_attr)
onnx_model_path = "debug_gnn_model.onnx"
# Compile the model
print("Compiling the model")
print("batch.x", batch.x)
start_compile = time.time()
quantized_numpy_module = compile_brevitas_qat_model(
wrapped_model, # Our model
batch.x, # A representative input-set to be used for both quantization and compilation\
output_onnx_file=onnx_model_path,
verbose=True
)
end_compile = time.time()
print(f"Compilation finished in {end_compile - start_compile:.2f} seconds")
# Check that the network is compatible with FHE constraints
bitwidth = quantized_numpy_module.fhe_circuit.graph.maximum_integer_bit_width()
print(
f"Max bit-width: {bitwidth} bits" + " -> it works in FHE!!"
if bitwidth <= 16
else " too high for FHE computation"
)
# Execute prediction using simulation
# (not encrypted but fast, and results are equivalent)
if not use_sim:
print("Generating key")
start_key = time.time()
quantized_numpy_module.fhe_circuit.keygen()
end_key = time.time()
print(f"Key generation finished in {end_key - start_key:.2f} seconds")
fhe_mode = "simulate" if use_sim else "execute"
predictions = np.zeros_like(y_test)
preds = []
ground_truths = []
for batch in tqdm.tqdm(te_loader, disable=not args.tqdm):
#select the seed edges from which the batch was created
inds = inds.detach().cpu()
batch_edge_inds = inds[batch.input_id.detach().cpu()]
batch_edge_ids = te_loader.data.edge_attr.detach().cpu()[batch_edge_inds, 0]
mask = torch.isin(batch.edge_attr[:, 0].detach().cpu(), batch_edge_ids)
#add the seed edges that have not been sampled to the batch
missing = ~torch.isin(batch_edge_ids, batch.edge_attr[:, 0].detach().cpu())
if missing.sum() != 0 and (args.data == 'Small_J' or args.data == 'Small_Q'):
missing_ids = batch_edge_ids[missing].int()
n_ids = batch.n_id
add_edge_index = te_data.edge_index[:, missing_ids].detach().clone()
node_mapping = {value.item(): idx for idx, value in enumerate(n_ids)}
add_edge_index = torch.tensor([[node_mapping[val.item()] for val in row] for row in add_edge_index])
add_edge_attr = te_data.edge_attr[missing_ids, :].detach().clone()
add_y = te_data.y[missing_ids].detach().clone()
batch.edge_index = torch.cat((batch.edge_index, add_edge_index), 1)
batch.edge_attr = torch.cat((batch.edge_attr, add_edge_attr), 0)
batch.y = torch.cat((batch.y, add_y), 0)
mask = torch.cat((mask, torch.ones(add_y.shape[0], dtype=torch.bool)))
#remove the unique edge id from the edge features, as it's no longer needed
batch.edge_attr = batch.edge_attr[:, 1:]
print("Starting inference")
start_infer = time.time()
pred = quantized_numpy_module.forward(batch.x).argmax(1)
end_infer = time.time()
print(f"Compilation finished in {end_compile - start_compile:.2f} seconds")
if not use_sim:
print(f"Key generation finished in {end_key - start_key:.2f} seconds")
print(
f"Inferences finished in {end_infer - start_infer:.2f} seconds "
f"({(end_infer - start_infer)/len(batch.x):.2f} seconds/sample)"
)
preds.append(pred)
ground_truths.append(batch.y[mask])
# Compute accuracy
pred = torch.cat(preds, dim=0).cpu().numpy()
ground_truth = torch.cat(ground_truths, dim=0).cpu().numpy()
accuracy = np.mean(pred == ground_truth) * 100
print(f"Test Quantized Accuracy: {accuracy:.2f}% on {len(X_test)} examples.")
f1 = f1_score(ground_truth, pred)
return bitwidth, accuracy, f1, predictions, quantized_numpy_module
def train_homo(tr_loader, val_loader, te_loader, tr_inds, val_inds, te_inds, model, optimizer, loss_fn, args, config, device, tr_data, val_data, te_data):
best_te_f1 = 0
total_training_time = 0
total_val_time = 0
total_test_time = 0
for epoch in range(config.epochs):
logging.info(f"\nEpoch {epoch}")
#training
train_start_time = time.time()
total_loss = total_examples = 0
preds = []
ground_truths = []
for batch in tqdm.tqdm(tr_loader, disable=not args.tqdm):
optimizer.zero_grad()
#select the seed edges from which the batch was created
inds = tr_inds.detach().cpu()
batch_edge_inds = inds[batch.input_id.detach().cpu()]
batch_edge_ids = tr_loader.data.edge_attr.detach().cpu()[batch_edge_inds, 0]
mask = torch.isin(batch.edge_attr[:, 0].detach().cpu(), batch_edge_ids)
#remove the unique edge id from the edge features, as it's no longer needed
batch.edge_attr = batch.edge_attr[:, 1:]
batch.to(device)
out = model(batch.x, batch.edge_index, batch.edge_attr)
pred = out[mask]
ground_truth = batch.y[mask]
preds.append(pred.argmax(dim=-1))
ground_truths.append(ground_truth)
loss = loss_fn(pred, ground_truth)
loss.backward()
optimizer.step()
total_loss += float(loss) * pred.numel()
total_examples += pred.numel()
pred = torch.cat(preds, dim=0).detach().cpu().numpy()
ground_truth = torch.cat(ground_truths, dim=0).detach().cpu().numpy()
f1 = f1_score(ground_truth, pred)
wandb.log({"f1/train": f1}, step=epoch)
logging.info(f'Train F1: {f1:.4f}')
train_time = time.time() - train_start_time
total_training_time += train_time
wandb.log({"time/train": train_time}, step=epoch)
logging.info(f'Training time: {train_time:.2f}s')
#evaluate
val_start_time = time.time()
val_f1 = evaluate_homo(val_loader, val_inds, model, val_data, device, args)
val_time = time.time() - val_start_time
total_val_time += val_time
test_start_time = time.time()
if args.fhe:
_, _, te_f1, clear_prediction, vl_quantized_numpy_module = compile_and_test(
tr_loader, te_loader, te_inds, model, tr_data.x, te_data, args, use_sim=True
)
else:
te_f1 = evaluate_homo(te_loader, te_inds, model, te_data, device, args)
test_time = time.time() - test_start_time
total_test_time += test_time
wandb.log({"time/val": val_time}, step=epoch)
wandb.log({"f1/validation": val_f1}, step=epoch)
wandb.log({"time/test": test_time}, step=epoch)
wandb.log({"f1/test": te_f1}, step=epoch)
logging.info(f'Validation time: {val_time:.2f}s')
logging.info(f'Validation F1: {val_f1:.4f}')
logging.info(f'Test time: {test_time:.2f}s')
logging.info(f'Test F1: {te_f1:.4f}')
if epoch == 0:
wandb.log({"best_test_f1": te_f1}, step=epoch)
elif te_f1 > best_te_f1:
best_te_f1 = te_f1
wandb.log({"best_test_f1": te_f1}, step=epoch)
if args.save_model:
save_model(model, optimizer, epoch, args)
if epoch == config.epochs-1: #if last epoch
mean_training_time = total_training_time/epoch
mean_val_time = total_val_time/epoch
mean_test_time = total_test_time/epoch
logging.info(f'Best Test F1: {te_f1:.5f}')
logging.info(f'Mean Training Time per epoch: {mean_training_time:.5f}s')
wandb.run.summary["time/mean_training_time"] = mean_training_time
logging.info(f'Mean Validation Time per epoch: {mean_val_time:.5f}s')
wandb.run.summary["time/mean_val_time"] = mean_val_time
logging.info(f'Mean Test Time per epoch: {mean_test_time:.5f}s')
wandb.run.summary["time/mean_test_time"] = mean_test_time
return model
def train_hetero(tr_loader, val_loader, te_loader, tr_inds, val_inds, te_inds, model, optimizer, loss_fn, args, config, device, val_data, te_data):
#training
best_val_f1 = 0
for epoch in range(config.epochs):
total_loss = total_examples = 0
preds = []
ground_truths = []
for batch in tqdm.tqdm(tr_loader, disable=not args.tqdm):
optimizer.zero_grad()
#select the seed edges from which the batch was created
inds = tr_inds.detach().cpu()
batch_edge_inds = inds[batch['node', 'to', 'node'].input_id.detach().cpu()]
batch_edge_ids = tr_loader.data['node', 'to', 'node'].edge_attr.detach().cpu()[batch_edge_inds, 0]
mask = torch.isin(batch['node', 'to', 'node'].edge_attr[:, 0].detach().cpu(), batch_edge_ids)
#remove the unique edge id from the edge features, as it's no longer needed
batch['node', 'to', 'node'].edge_attr = batch['node', 'to', 'node'].edge_attr[:, 1:]
batch['node', 'rev_to', 'node'].edge_attr = batch['node', 'rev_to', 'node'].edge_attr[:, 1:]
batch.to(device)
out = model(batch.x_dict, batch.edge_index_dict, batch.edge_attr_dict)
out = out[('node', 'to', 'node')]
pred = out[mask]
ground_truth = batch['node', 'to', 'node'].y[mask]
preds.append(pred.argmax(dim=-1))
ground_truths.append(batch['node', 'to', 'node'].y[mask])
loss = loss_fn(pred, ground_truth)
loss.backward()
optimizer.step()
total_loss += float(loss) * pred.numel()
total_examples += pred.numel()
pred = torch.cat(preds, dim=0).detach().cpu().numpy()
ground_truth = torch.cat(ground_truths, dim=0).detach().cpu().numpy()
f1 = f1_score(ground_truth, pred)
wandb.log({"f1/train": f1}, step=epoch)
logging.info(f'Train F1: {f1:.4f}')
#evaluate
val_f1 = evaluate_hetero(val_loader, val_inds, model, val_data, device, args)
te_f1 = evaluate_hetero(te_loader, te_inds, model, te_data, device, args)
wandb.log({"f1/validation": val_f1}, step=epoch)
wandb.log({"f1/test": te_f1}, step=epoch)
logging.info(f'Validation F1: {val_f1:.4f}')
logging.info(f'Test F1: {te_f1:.4f}')
if epoch == 0:
wandb.log({"best_test_f1": te_f1}, step=epoch)
elif val_f1 > best_val_f1:
best_val_f1 = val_f1
wandb.log({"best_test_f1": te_f1}, step=epoch)
if args.save_model:
save_model(model, optimizer, epoch, args)
return model
def get_model(sample_batch, config, args):
n_feats = sample_batch.x.shape[1] if not isinstance(sample_batch, HeteroData) else sample_batch['node'].x.shape[1]
e_dim = (sample_batch.edge_attr.shape[1] - 1) if not isinstance(sample_batch, HeteroData) else (sample_batch['node', 'to', 'node'].edge_attr.shape[1] - 1)
if args.model == "gin":
model = GINe(
num_features=n_feats, num_gnn_layers=config.n_gnn_layers, n_classes=2,
n_hidden=round(config.n_hidden), residual=False, edge_updates=args.emlps, edge_dim=e_dim,
dropout=config.dropout, final_dropout=config.final_dropout
)
elif args.model == "gin_fhe":
model = GINe_FHE(
num_features=n_feats, num_gnn_layers=config.n_gnn_layers, n_classes=2,
n_hidden=round(config.n_hidden), residual=False, edge_updates=args.emlps, edge_dim=e_dim,
dropout=config.dropout, final_dropout=config.final_dropout
)
elif args.model == "gat":
model = GATe(
num_features=n_feats, num_gnn_layers=config.n_gnn_layers, n_classes=2,
n_hidden=round(config.n_hidden), n_heads=round(config.n_heads),
edge_updates=args.emlps, edge_dim=e_dim,
dropout=config.dropout, final_dropout=config.final_dropout
)
elif args.model == "pna":
if not isinstance(sample_batch, HeteroData):
d = degree(sample_batch.edge_index[1], dtype=torch.long)
else:
index = torch.cat((sample_batch['node', 'to', 'node'].edge_index[1], sample_batch['node', 'rev_to', 'node'].edge_index[1]), 0)
d = degree(index, dtype=torch.long)
deg = torch.bincount(d, minlength=1)
model = PNA(
num_features=n_feats, num_gnn_layers=config.n_gnn_layers, n_classes=2,
n_hidden=round(config.n_hidden), edge_updates=args.emlps, edge_dim=e_dim,
dropout=config.dropout, deg=deg, final_dropout=config.final_dropout
)
elif config.model == "rgcn":
model = RGCN(
num_features=n_feats, edge_dim=e_dim, num_relations=8, num_gnn_layers=round(config.n_gnn_layers),
n_classes=2, n_hidden=round(config.n_hidden),
edge_update=args.emlps, dropout=config.dropout, final_dropout=config.final_dropout, n_bases=None #(maybe)
)
return model
def train_gnn(tr_data, val_data, te_data, tr_inds, val_inds, te_inds, args):
#set device
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
#define a model config dictionary and wandb logging at the same time
wandb.init(
mode="disabled" if args.testing else "online",
project="your_proj_name", #replace this with your wandb project name if you want to use wandb logging
config={
"epochs": args.n_epochs,
"batch_size": args.batch_size,
"model": args.model,
"data": args.data,
"num_neighbors": args.num_neighs,
"lr": extract_param("lr", args),
"n_hidden": extract_param("n_hidden", args),
"n_gnn_layers": extract_param("n_gnn_layers", args),
"loss": "ce",
"w_ce1": extract_param("w_ce1", args),
"w_ce2": extract_param("w_ce2", args),
"dropout": extract_param("dropout", args),
"final_dropout": extract_param("final_dropout", args),
"n_heads": extract_param("n_heads", args) if args.model == 'gat' else None
}
)
config = wandb.config
#set the transform if ego ids should be used
if args.ego:
transform = AddEgoIds()
else:
transform = None
#add the unique ids to later find the seed edges
add_arange_ids([tr_data, val_data, te_data])
tr_loader, val_loader, te_loader = get_loaders(tr_data, val_data, te_data, tr_inds, val_inds, te_inds, transform, args)
#get the model
sample_batch = next(iter(tr_loader))
model = get_model(sample_batch, config, args)
if args.finetune:
model, optimizer = load_model(model, device, args, config)
else:
model = get_model(sample_batch, config, args)
model.to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=config.lr)
if args.reverse_mp:
model = to_hetero(model, te_data.metadata(), aggr='mean')
sample_x = sample_batch.x if not isinstance(sample_batch, HeteroData) else sample_batch.x_dict
sample_edge_index = sample_batch.edge_index if not isinstance(sample_batch, HeteroData) else sample_batch.edge_index_dict
if isinstance(sample_batch, HeteroData):
sample_batch['node', 'to', 'node'].edge_attr = sample_batch['node', 'to', 'node'].edge_attr[:, 1:]
sample_batch['node', 'rev_to', 'node'].edge_attr = sample_batch['node', 'rev_to', 'node'].edge_attr[:, 1:]
else:
sample_batch.edge_attr = sample_batch.edge_attr[:, 1:]
sample_edge_attr = sample_batch.edge_attr if not isinstance(sample_batch, HeteroData) else sample_batch.edge_attr_dict
logging.info(summary(model, sample_x, sample_edge_index, sample_edge_attr))
loss_fn = torch.nn.CrossEntropyLoss(weight=torch.FloatTensor([config.w_ce1, config.w_ce2]).to(device))
if args.reverse_mp:
model = train_hetero(tr_loader, val_loader, te_loader, tr_inds, val_inds, te_inds, model, optimizer, loss_fn, args, config, device, val_data, te_data)
else:
model = train_homo(tr_loader, val_loader, te_loader, tr_inds, val_inds, te_inds, model, optimizer, loss_fn, args, config, device, tr_data, val_data, te_data)
wandb.finish()